MARC CLUSTER BENCHMARK

1
MARC CLUSTER BENCHMARK

• Non-Linear Analysis
• of a
• Suspension Knuckle
• using
• MSC Marc 2003 DDM

2
HP J6700 1 CPU Solve Time 1326 Baseline Time
3
Linux Cluster 2 CPUs Solve Time 520 Speed-Up
2.5X
4
Linux Cluster 4 CPUs Solve Time 307 Speed-Up
4.3X
5
Linux Cluster 8 CPUs Solve Time 151 Speed-Up
7.3X
6
Linux Cluster 16 CPUs Solve Time 103 Speed-Up
12.8X
7
(No Transcript)
8
(No Transcript)
9
Benchmark Summary

• For the current benchmark part, it would appear
  that 4 CPUs provides optimal solution time vs
  computing resources being utilized. However, it
  should be expected that different types of
  non-linear simulations will yield varying
  performance curves and require different numbers
  of CPUs due to the diverse nature of non-linear
  problems.
• While not difficult, submission of Marc
  simulations to the cluster is not currently as
  simple as pushing go. There are a number of
  variables that have to be configured to ensure a
  successful run that yields results rather than
  errors. We are working on a more efficient way to
  accomplish this task. A cookbook approach if
  you will
• Irregardless of the number of processors or the
  computing platform used, the resultant stress
  value was always the same (274 Mpa) and in the
  same location. No deviation in accuracy was
  detected.
• There seems to currently be a bug in the Linux
  version of the Mentat Pre-processing software.
  The HP version had to be used to create the input
  decks for the cluster runs. MSC has been alerted
  to this and is working on correcting the issue.
• The Linux Networx hardware has so far performed
  flawlessly in all of the categories that it has
  been tested against. They have also been very
  professional in lending support to MSC to help
  resolve software issues in the Marc application.

10
(No Transcript)
11
(No Transcript)